> On Apr 19, 2022, at 6:09 PM, Trond Myklebust <trondmy@xxxxxxxxxxxxxxx> wrote:
>
> On Tue, 2022-04-19 at 19:40 +0000, Chuck Lever III wrote:
>>
>>
>>> On Apr 19, 2022, at 3:04 PM, Trond Myklebust
>>> <trondmy@xxxxxxxxxxxxxxx> wrote:
>>>
>>> On Tue, 2022-04-19 at 18:16 +0000, Chuck Lever III wrote:
>>>>
>>>>
>>>>> On Apr 18, 2022, at 10:57 PM, Trond Myklebust
>>>>> <trondmy@xxxxxxxxxxxxxxx> wrote:
>>>>>
>>>>>> On Mon, 2022-04-18 at 12:52 -0400, Chuck Lever wrote:
>>>>>>> Introduce a mechanism to cause xprt_transmit() to break out
>>>>>>> of
>>>>>>> its
>>>>>>> sending loop at a specific rpc_rqst, rather than draining the
>>>>>>> whole
>>>>>>> transmit queue.
>>>>>>>
>>>>>>> This enables the client to send just an RPC TLS probe and
>>>>>>> then
>>>>>>> wait
>>>>>>> for the response before proceeding with the rest of the
>>>>>>> queue.
>>>>>>>
>>>>>>> Signed-off-by: Chuck Lever <chuck.lever@xxxxxxxxxx>
>>>>>>> ---
>>>>>>> include/linux/sunrpc/sched.h | 2 ++
>>>>>>> include/trace/events/sunrpc.h | 1 +
>>>>>>> net/sunrpc/xprt.c | 2 ++
>>>>>>> 3 files changed, 5 insertions(+)
>>>>>>>
>>>>>>> diff --git a/include/linux/sunrpc/sched.h
>>>>>>> b/include/linux/sunrpc/sched.h
>>>>>>> index 599133fb3c63..f8c09638fa69 100644
>>>>>>> --- a/include/linux/sunrpc/sched.h
>>>>>>> +++ b/include/linux/sunrpc/sched.h
>>>>>>> @@ -125,6 +125,7 @@ struct rpc_task_setup {
>>>>>>> #define RPC_TASK_TLSCRED 0x00000008 /*
>>>>>>> Use
>>>>>>> AUTH_TLS credential */
>>>>>>> #define RPC_TASK_NULLCREDS 0x00000010 /*
>>>>>>> Use
>>>>>>> AUTH_NULL credential */
>>>>>>> #define RPC_CALL_MAJORSEEN 0x00000020 /*
>>>>>>> major
>>>>>>> timeout seen */
>>>>>>> +#define RPC_TASK_CORK 0x00000040 /*
>>>>>>> cork
>>>>>>> the
>>>>>>> xmit queue */
>>>>>>> #define RPC_TASK_DYNAMIC 0x00000080 /*
>>>>>>> task
>>>>>>> was
>>>>>>> kmalloc'ed */
>>>>>>> #define RPC_TASK_NO_ROUND_ROBIN
>>>>>>> 0x00000100
>>>>>>> /*
>>>>>>> send requests on "main" xprt */
>>>>>>> #define RPC_TASK_SOFT 0x00000200 /*
>>>>>>> Use
>>>>>>> soft
>>>>>>> timeouts */
>>>>>>> @@ -137,6 +138,7 @@ struct rpc_task_setup {
>>>>>>>
>>>>>>> #define RPC_IS_ASYNC(t) ((t)->tk_flags &
>>>>>>> RPC_TASK_ASYNC)
>>>>>>> #define RPC_IS_SWAPPER(t) ((t)->tk_flags &
>>>>>>> RPC_TASK_SWAPPER)
>>>>>>> +#define RPC_IS_CORK(t) ((t)->tk_flags &
>>>>>>> RPC_TASK_CORK)
>>>>>>> #define RPC_IS_SOFT(t) ((t)->tk_flags &
>>>>>>> (RPC_TASK_SOFT|RPC_TASK_TIMEOUT))
>>>>>>> #define RPC_IS_SOFTCONN(t) ((t)->tk_flags &
>>>>>>> RPC_TASK_SOFTCONN)
>>>>>>> #define RPC_WAS_SENT(t) ((t)->tk_flags &
>>>>>>> RPC_TASK_SENT)
>>>>>>> diff --git a/include/trace/events/sunrpc.h
>>>>>>> b/include/trace/events/sunrpc.h
>>>>>>> index 811187c47ebb..e8d6adff1a50 100644
>>>>>>> --- a/include/trace/events/sunrpc.h
>>>>>>> +++ b/include/trace/events/sunrpc.h
>>>>>>> @@ -312,6 +312,7 @@ TRACE_EVENT(rpc_request,
>>>>>>> { RPC_TASK_TLSCRED, "TLSCRED"
>>>>>>> }, \
>>>>>>> { RPC_TASK_NULLCREDS, "NULLCREDS"
>>>>>>> }, \
>>>>>>> { RPC_CALL_MAJORSEEN, "MAJORSEEN"
>>>>>>> }, \
>>>>>>> + { RPC_TASK_CORK, "CORK"
>>>>>>> }, \
>>>>>>> { RPC_TASK_DYNAMIC, "DYNAMIC"
>>>>>>> }, \
>>>>>>> { RPC_TASK_NO_ROUND_ROBIN, "NO_ROUND_ROBIN"
>>>>>>> }, \
>>>>>>> { RPC_TASK_SOFT, "SOFT"
>>>>>>> }, \
>>>>>>> diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
>>>>>>> index 86d62cffba0d..4b303b945b51 100644
>>>>>>> --- a/net/sunrpc/xprt.c
>>>>>>> +++ b/net/sunrpc/xprt.c
>>>>>>> @@ -1622,6 +1622,8 @@ xprt_transmit(struct rpc_task *task)
>>>>>>> if (xprt_request_data_received(task) &&
>>>>>>> !test_bit(RPC_TASK_NEED_XMIT, &task-
>>>>>>>> tk_runstate))
>>>>>>> break;
>>>>>>> + if (RPC_IS_CORK(task))
>>>>>>> + break;
>>>>>>> cond_resched_lock(&xprt->queue_lock);
>>>>>>> }
>>>>>>> spin_unlock(&xprt->queue_lock);
>>>>>>>
>>>>>>>
>>>>>>
>>>>>> This is entirely the wrong place for this kind of control
>>>>>> mechanism.
>>>>>
>>>>> I'm not sure I entirely understand your concern, so bear with
>>>>> me while I try to clarify.
>>>>>
>>>>>
>>>>>> TLS vs not-TLS needs to be decided up front when we initialise
>>>>>> the
>>>>>> transport (i.e. at mount time or whenever the pNFS channels are
>>>>>> set
>>>>>> up). Otherwise, we're vulnerable to downgrade attacks.
>>>>>
>>>>> Downgrade attacks are prevented by using "xprtsec=tls" because
>>>>> in that case, transport creation fails if either the AUTH_TLS
>>>>> fails or the handshake fails.
>>>>>
>>>>> The TCP connection has to be established first, though. Then the
>>>>> client can send the RPC_AUTH_TLS probe, which is the same as the
>>>>> NULL ping that it already sends. That mechanism is independent
>>>>> of the lower layer transport (TCP in this case).
>>>>>
>>>>> Therefore, RPC traffic must be stoppered while the client:
>>>>>
>>>>> 1. waits for the AUTH_TLS probe's reply, and
>>>>>
>>>>> 2. waits for the handshake to complete
>>>>>
>>>>> Because an RPC message is involved in this interaction, I didn't
>>>>> see a way to implement it completely within xprtsock's TCP
>>>>> connection logic. IMO, driving the handshake has to be done by
>>>>> the generic RPC client.
>>>>>
>>>>> So, do you mean that I need to replace RPC_TASK_CORK with a
>>>>> special return code from xs_tcp_send_request() ?
>>>
>>>
>>> I mean the right mechanism for controlling whether or not the
>>> transport
>>> is ready to serve RPC requests is through the XPRT_CONNECTED flag.
>>> All
>>> the existing generic RPC error handling, congestion handling, etc
>>> depends on that flag being set correctly.
>>>
>>> Until the TLS socket has completed its handshake protocol and is
>>> ready
>>> to transmit data, it should not be declared connected. The
>>> distinction
>>> between the two states 'TCP is unconnected' and 'TLS handshake is
>>> incomplete' is a socket/transport setup detail as far as the RPC
>>> xprt
>>> layer is concerned: just another set of intermediate states between
>>> SYN_SENT and ESTABLISHED.
>>
>> First, TLS is technically an upper layer protocol. It's not
>> part of the transport protocol. This is exactly how it's
>> implemented in the Linux kernel. And, TLS works on transports
>> other than TCP, so that makes it a reasonable candidate for
>> treatment in the generic client rather than in a particular
>> transport mechanism.
>
> Sorry, but no! As far as the RPC layer is concerned, there is no
> difference between a TLS socket and a TCP socket. The xprt layer should
> not have to know or care about the existence of TLS other that as a
> transport option to be configured at connection time.
>
>>
>> Second, the "intermediate states" would be /outside/ of SYN_SENT
>> and ESTABLISHED. A TCP transport has to be in the ESTABLISHED
>> state (ie, the transport's connection handshake has to be
>> complete) before any TLS traffic can go over it.
>>
>
> My point is we don't give a damn about the intermediate states in the
> RPC layer.
>
>> Most importantly, the client has to send an RPC message first
>> before it can start a TLS handshake. The RPC-with-TLS protocol
>> specification requires that the handshake be preceded with the
>> NULL AUTH_TLS request, which is an RPC. Otherwise, there's no
>> way for the server end to know when to expect a handshake.
>>
>
> Sure, but those are 2 non-overlapping states. The socket is first in a
> state where it needs to do a NULL ping using regular RPC/TCP. Then it
> needs to do the TLS handshake. Then it transitions into the state where
> it can act like any other transport.
Understood: this architecture more-or-less mimics what the RPC client would see for a transport like QUIC where connection establishment and the security handshake are integrated and handled concurrently.
The reality is that the RPC client’s transport layer will have to deal with these steps separately for TLS-on-UDP and TLS-on-TCP, but hiding the details under the transport switch is fine with me as long as there is a way to send the AUTH_TLS probe before the transport is marked “connected” (see below).
>> In today's RPC client, the underlying connection has to be in
>> the XPRT_CONNECTED state before the RPC client can exchange any
>> RPC transaction, including AUTH_TLS NULL.
>>
>> To make it work the way you've suggested, we would have to build
>> a mechanism that could send the AUTH_TLS NULL and receive and
>> parse its reply /before/ the client has put the transport into
>> the XPRT_CONNECTED state, and that NULL request would have to
>> be driven from inside the transport instance (not via the FSM
>> where all other RPC traffic originates).
>>
>> Do you have any suggestions about how to make this last point
>> less painful?
>
> This isn't too different from what we already do with the rpcbind call
> for performing port discovery. The only difference is that the NULL
> ping needs to happen on the same transport as the one being constructed
> and that it needs to happen after the TCP connection is complete.
>
> So I'd suggest that TLS/TCP needs to be a different xprt_class than the
> base TCP, then doing the whole "do-NULL-ping-and-TLS-handshake" in the
> connect() callback for that new class.
>
> The connect() callback can set up a private rpc client and do the NULL
> call asynchronously just like we do in rpcb_getport_async().
“Set up a separate rpc_client to handle the AUTH_TLS probe” was the piece I was missing. The rest of this is pretty much what the RPC changes in this patch series already do, but organized a little differently (for example, they use a “done” callback as you describe below, so we’re already about halfway there).
The idea is to stopper the stream of RPC messages by leaving the xprt marked “not connected” instead of by adding and setting an RPC_TASK_CORK flag. The FSM will no longer be involved at all in dealing with TLS.
> When the
> RPC call completes, we steal the resulting socket from that private rpc
> client and kick off the TLS handshake on it. All that can be done in
> the rpc_call_done callback (i.e. the equivalent of rpcb_getport_done).
>
> Once the TLS handshake is done, you can set the XPRT_CONNECTED state
> and call xprt_wake_pending_tasks().
